Where 195 B2B GTM teams sit on the curve, and what moves them to the next stage.
47% of B2B go-to-market teams have zero AI agents in production. Only 8% have more than ten. Only 2% have more than twenty.
That distribution comes from the 2026 State of AI for GTM report by Kyle Poyar (Growth Unhinged) and Maja Voje (GTM Strategist), published in January and surveying 195 GTM leaders. Their headline framing is sharp. We have entered the “have and have not era” of AI in GTM. 53% of leaders see no impact from AI. 24% see big impact.
The split is real. But have/have not is binary, and the data tells a different story. The agent-count distribution is a curve with five distinct stages. Each stage has a dominant behavior, a bottleneck, and a specific move that gets a team to the next one. Most teams do not get stuck because they lack tools. They get stuck because they do not recognize which stage they are in, and they keep trying to solve the wrong problem.
Here is the curve.
What does it look like to have zero AI agents in GTM?
Stage 0. Watching. 47% of teams.
A few licensed AI tools, vendor-controlled. Maybe a Custom GPT someone built once and shared in Slack. The team treats AI as something other people do.
The bottleneck at this stage is the absence of an internal proof point. Nobody on the team has shipped one. Risk-averse leadership has nothing to point at when a vendor pitch lands. The team buys more tools instead of building any agents, because tools feel safer than systems.
The move out of stage 0 is one small agent that produces a measurable outcome. It does not have to touch revenue. It just has to do something the team could not do before, with a number attached. One agent with one number is enough to put the next budget conversation in a different posture.
Watch for tool sprawl as a substitute for system building. Buying a fifth AI vendor is not progress.
What is the first stage of AI agent adoption?
Stage 1. Experimenting. 32% of teams.
One to three agents in production, usually narrow in scope. One for content. One for research. One for outbound enrichment. Built by one person on the team, often the most curious operator.
The bottleneck is single point of failure. The person who built the agents is also the only person who runs them. There is no shared playbook. If they leave, the agents die. If they take a week off, the agents drift.
The move out of stage 1 is documentation. Light evals. The cultural shift from “my agent” to “our agent.” This is also where context layers start to matter. Shared prompts, CLAUDE.md files, MCP-style tool definitions. The point is not engineering elegance. The point is that the next person on the team can run the agent without calling the builder.
Watch for agents that work in demo but break in production because nobody traced what they actually do.
What happens when a GTM team scales past three AI agents?
Stage 2. Stacking. 14% of teams.
Four to ten agents in production. Coverage spans content, prospecting, research, ops. Each agent solves a discrete job.
The bottleneck is that the agents do not talk to each other. Information passes through humans. Cost discipline becomes a real concern as the agents start running on schedules instead of on demand. This is the agent zoo problem. A collection of working agents that does not yet behave like a system.
Stage 2 breaks open with shared context. Agent-to-agent handoffs. Observability. Cost tracing. This is the stage where the underlying infrastructure starts to matter more than the individual agents. Memory. Eval framework. Orchestration.
Watch for the trap of building five disconnected agents and calling that a system. It is a collection.
What does production-grade AI agent operations look like?
Stage 3. Operating. 6% of teams.
Eleven to twenty agents. Agents are infrastructure now. They run on schedules. They are triggered by events. They do real GTM work without daily human supervision. The team treats them like coworkers and runs reviews on them.
The bottleneck is governance. Permissions. What can the agent do, what can it not do, who approves the edge cases. Cost is a board-level concern.
The move out of stage 3 is read-only by default. Scoped tokens. Tool-level permissions. Eval and regression rigor on every shipped change. Formal change management for agent updates. This is where physical enforcement stops being optional.
Watch for the confusion between instructions and constraints. If the agent can do it, the prompt will not stop it.
What does AI agent maturity look like at scale?
Stage 4. Compounding. 2% of teams.
Twenty-one or more agents. Agents build agents. New capabilities ship in days, not quarters. The team’s headcount has decoupled from the org’s output.
The bottleneck is talent. People who can run this layer at scale are rare. Recruiting and retention become the new constraint.
Stage 4 has a different shape than the stages below it. The work is no longer about building the next agent. It is about building the next generation of operators. Public artifacts so prospective hires self-select in. Hire on the operator-engineer axis, not pure engineer or pure GTM.
Watch for architecture debt. The systems that worked at ten agents do not work at thirty unless they were built that way from the start.
The Pattern Across All Five Stages
Each stage requires a different kind of work. Stage 0 needs a proof point. Stage 1 needs documentation. Stage 2 needs infrastructure. Stage 3 needs governance. Stage 4 needs the next generation. The mistake is thinking you can skip stages. The other mistake is thinking the stage 2 bottleneck is the same as the stage 4 bottleneck.
This is also why “buy more AI tools” does not move anyone up the curve. The tools are mostly the same across stages. 91% of GTM teams in the report use general-purpose models like Claude, ChatGPT, or Gemini. The differentiator at every stage is the system around the model, not the model itself.
So where does your team sit? What is the next move? Most teams over-index on the gap to the top. The stage above yours is the one to focus on.
The teams that compound are not the ones with the best models. They are the ones who recognized which stage they were in and built the move for the next one.
FAQ
What percentage of B2B GTM teams have AI agents in production?
According to the 2026 State of AI for GTM report by Kyle Poyar and Maja Voje, surveying 195 B2B GTM leaders, 47% have zero AI agents in production. 32% have one to three. 14% have four to ten. 6% have eleven to twenty. Only 2% have twenty-one or more.
What is the difference between using AI tools and running AI agents in GTM?
AI tools are vendor-controlled products that wrap a model in a narrow interface. AI agents are systems your team builds and runs, with their own context, tool access, and workflows. The 2026 State of AI for GTM report found that 91% of GTM teams use general-purpose AI tools like Claude or ChatGPT, but only 53% have any agents in production. Tools are the starting point. Agents are the system.
How many AI agents should a GTM team have?
There is no correct number. The right question is which stage your team is in and what the next move looks like. A team at stage 1 with two well-documented agents is in better shape than a team at stage 2 with seven brittle ones. The agents matter less than the system around them.
What is the bottleneck at each stage of AI agent adoption?
Stage 0 is the absence of an internal proof point. Stage 1 is single-point-of-failure dependency on the builder. Stage 2 is the lack of agent-to-agent infrastructure. Stage 3 is governance and permissions. Stage 4 is talent.
What moves a team from stage 1 to stage 2 of AI agent maturity?
Documentation. Light evals. Shared context layers like CLAUDE.md files and standardized prompts. The cultural shift from “my agent” to “our agent.” The goal is that the next person on the team can run the agent without calling the original builder.